2026-03-14 Image Agent Runtime and Chat Review Flow

This release upgrades the image agent from a simple chat-triggered generator to a job-based runtime with post-generation review, auto-revision, and user handoff when needed. It also fixes several chat UX regressions introduced while adding critic and review flows.

User-visible changes

• Images generated from the AI assistant now appear in chat immediately when the first result is ready.
• The task center now uses clearer states and no longer treats requires_action as a failed generation.
• Action cards now appear after the generated image instead of awkwardly appearing before the visual result.
• Success sounds play when the image is actually ready; failure sounds still only play for real failures.
• Leaked tool-planning JSON is removed from chat responses.
• Action-card copy now has safer Chinese/English fallback behavior.

Agent behavior changes

• Added structured image critic review with issue typing, quality signals, revision strength, and review traces for development.
• Reduced unnecessary interruptions by preferring auto_revise for low-risk, fixable issues.
• Expanded cases that should pause for confirmation, especially vague aesthetic/global requests, broad composition changes, and ambiguous direction shifts.
• Added calibration so the agent is less likely to over-edit already strong results.

Internal/runtime changes

• Image generation now uses a more explicit split between user-visible completion and background critic/review processing.
• Job and task states were refined to better distinguish generating, reviewing, action required, completed, and failed.

QA

• Regression checklist for action-card behavior is documented in docs/qa/image-agent-card-regression.md.